deepseek-v4.1POST
oMLX 0.7.2 boosts DeepSeek V4.1 decode 38% on Apple Silicon
oMLX 0.7.2 raises DeepSeek V4.1 decode from 49.4 to 68.2 tok/s on the oQ3e quant, with prefill holding near 900 tok/s up to a 128K prompt on an M3 Ultra 512GB.
Brad Mills
@BradMillsArchived 10 days agoOriginally posted 2026.10.21
Brad Mills @BradMillsOriginally posted 2026.10.21
oMLX 0.7.2 dropped a day after 0.7.1. Faster DeepSeek V4.1 on every Apple Silicon build, decode 49.4 to 68.2 tok/s on the oQ3e quant plus 38%. Prefill holds around 900 tok/s all the way to a 128K prompt, measured on an M3 Ultra 512GB. #LocalLLaMA
x.com/bradmills/status/2112897535517811066Archived 2026.10.22 12:01